Since replay-single 61d's debut in Jul, 2008, the test has been led astray from its original mission. The historical changes described below should be reverted or fixed.
llog_setup_named() and llog_setup(). Originally, the fault injection simulated what might happen when the "CATALOGS" internal file was corrupted. After the changes were applied, the fault started to be injected at the first call of llog_setup_named()/llog_setup() even when lop_setup was not implemented. In other words, a fault could be injected at locations that would never fail! Another issue in these changes is that replacing llog_ctxt_destroy() with llog_ctxt_put() causes lop_cleanup to be invoked upon lop_setup failures. In general, interfaces like lop_setup should clean things up before returning an error; lop_cleanup should not be called upon lop_setup failures.
test_61d(). This change looks like a result of the fictional fault injection described above. The test should stop and start the MDT rather than the MGS.
This was enabled in patch https://review.whamcloud.com/27404 "
LU-9590tests: remove replay-single tests from ALWAYS_EXCEPT".